A thesis on the application of neural network computing to the constrained flight control allocation problem
MetadataShow full item record
The feasibility of utilizing a neural network to solve the constrained flight control allocation problem is investigated for the purposes of developing guidelines for the selection of a neural network structure as a function of the control allocation problem parameters. The control allocation problem of finding the combination of several flight controls that generate a desired body axis moment without violating any control constraint is considered. Since the number of controls, which are assumed to be individually linear and constrained to specified ranges, is in general greater than the number of moments being controlled, the problem is nontrivial. Parallel investigations in direct and generalized inverse solutions have yielded a software tool (namely CAT, for Control Allocation Toolbox) to provide neural network training, testing, and comparison data. A modified back propagation neural network architecture is utilized to train a neural network to emulate the direct allocation scheme implemented in CAT, which is optimal in terms of having the ability to attain all possible moments with respect to a given control surface configuration. Experimentally verified heuristic arguments are employed to develop guidelines for the selection of neural network configuration and parameters with respect to a general control allocation problem. The control allocation problem is shown to be well suited for a neural network solution. Specifically, a six hidden neuron neural network is shown to have the ability to train efficiently, fonn an effective neural network representation of the subset of attainable moments, and independently discover the internal relationships between moments and controls. The performance of the neural network control allocator, trained on the basis of the developed guidelines, is examined for the reallocation of a seven control surface configuration representative of the F/A-18 HARV in a test maneuver flown using the original control laws of an existing flight simulator. The trained neural network is found to have good overall generalization performance, although limitations arise from the ability to obtain the resolution of the direct allocation scheme at low moment requirements. Lastly, recommendations offered include: ( 1) a proposed application to other unwieldy control al1ocation algorithms, with possible accounting for control actuator rate limitations, so that the computational superiority of the neural network could be fully realized; and (2) the exploitation of the adaptive aspects of neural network computing.
- Masters Theses